US, Taiwan and EU take different paths on AI medical devices

The US Food and Drug Administration (FDA) revised two digital-health policies in January 2026, drawing separate boundaries for low-risk wellness products and clinical decision-support software. Taiwan's framework sends qualifying AI medical software through product registration, while the European Union combines medical-device rules with a second set of obligations for high-risk AI.

These systems cannot be ranked on a single scale from permissive to strict. They decide different questions at different points: whether software is a medical device, what evidence is reviewed before sale, how planned model changes are controlled and which duties continue after a product reaches the market.

The US boundary starts with intended use

The FDA's January 6 general-wellness guidance addresses products intended to encourage a healthy lifestyle. US law excludes certain software functions from the device definition when they support a healthy lifestyle and are unrelated to diagnosing, curing, mitigating, preventing or treating a disease. For low-risk products that may still meet the device definition, the guidance describes circumstances in which the agency does not intend to examine compliance with device requirements.

That policy is narrower than a blanket exemption for wearables. Intended use and promotional claims matter. A product presented as tracking general activity does not occupy the same regulatory position as software marketed to diagnose a named condition or direct treatment.

The FDA issued separate final guidance for clinical decision-support software on January 29. The document explains the statutory criteria that can exclude some decision-support functions from the US device definition, while software used by patients or caregivers remains subject to existing digital-health policies when it meets that definition. The result is a function-by-function boundary, not a rule that all AI-assisted recommendations avoid review.

Products on the other side of that boundary still enter established US pathways. The FDA says products on its AI-enabled device list met applicable premarket requirements, including a focused review of safety and effectiveness for the intended use. The agency also warns that the list is not comprehensive because it identifies devices mainly through AI-related terms in public authorization records.

Taiwan places AI software inside device registration

Taiwan's Food and Drug Administration, the health ministry agency responsible for medical-product oversight, publishes a dedicated registration guide for software that uses artificial intelligence or machine learning as a medical device. Its English-language guidance says qualifying software must comply with Taiwan's Medical Devices Act and describes registration material covering the product, algorithm, training data, performance evaluation, cybersecurity and risk management.

The guide does not make every algorithm used in health care a medical device. Its scope is software that uses clinical data to perform a medical-device function, including products in which AI supplies only part of that function. Classification and registration then follow Taiwan's device framework rather than a separate AI authorization system.

This creates a different starting point from the US policy documents. The US guidance examined here devotes substantial attention to functions that fall outside the device definition or receive enforcement discretion. Taiwan's AI software guide is written for manufacturers whose products are already within the medical-device registration process.

The comparison does not establish which authority reviews faster or which route produces better clinical outcomes. APPI News could not find directly comparable official data on review times, rejection rates or post-market enforcement outcomes across the three markets. Product counts also cannot answer that question because regulators publish lists with different scopes and update schedules.

The EU adds AI duties to product regulation

AI medical devices sold in the European Union may have to comply with both the Medical Device Regulation or In Vitro Diagnostic Medical Devices Regulation and the AI Act. Article 6 of the AI Act classifies an AI system as high-risk when it is a product or safety component covered by listed EU legislation and that product requires a third-party conformity assessment. The two conditions mean the label “medical AI” alone does not determine classification.

For systems that qualify, the AI Act adds requirements covering risk management, data governance, technical documentation, logs, transparency, human oversight, accuracy, robustness and cybersecurity. The European Commission says providers must conduct a conformity assessment, maintain a quality-management system and remain responsible for safety and compliance throughout the system's lifecycle.

The timing changed after the source report was published. The European Commission now says the high-risk rules for AI embedded in regulated products, including medical devices, will apply from August 2, 2028. The extension followed delays in standards and implementation support, leaving manufacturers with more preparation time but also a moving compliance schedule.

Three frameworks shift risk to different checkpoints

The US materials create regulatory space before device review by excluding specified functions or describing enforcement discretion for low-risk products. That approach depends heavily on intended-use statements, product design and the agency's ability to respond when claims or functions cross the line. Enforcement discretion is an agency policy on applying specified requirements, not marketing authorization.

Taiwan's published guide concentrates on the evidence package for AI software already treated as a medical device. The EU route can place AI-specific governance duties on top of sectoral conformity assessment. In practice, a developer seeking access to all three markets needs separate classification and evidence strategies rather than one universal approval file.

The central policy choice is where uncertainty is handled. A broad premarket boundary puts more questions into formal review; a narrow boundary leaves more products outside that gate and increases the importance of surveillance and enforcement. Layering AI governance onto device law adds documentation and oversight before and after market entry, but the effect cannot be measured from the text of the rules alone.

Frequently asked questions

Did the FDA exempt all AI wearables in 2026?
No. The general-wellness policy covers low-risk products and turns on intended use and claims. Software marketed for diagnosis or treatment can still meet the US device definition and face applicable requirements.

Does the FDA list show every AI medical device authorized in the United States?
No. The FDA describes the list as a transparency resource and explicitly says it is not comprehensive. Devices may be added later when public decision summaries become available.

Does Taiwan have a separate AI medical-device law?
The cited Taiwan guidance places qualifying AI and machine-learning software within the territory's Medical Devices Act and registration system. It supplies technical expectations for AI software rather than creating a separate authorization outside device law.

Are all AI medical devices high-risk under the EU AI Act?
No. The product-based route requires both coverage by listed EU product legislation and a third-party conformity assessment. The product's classification and intended purpose determine whether those conditions are met.